Hybrid Approach to Content Recommendation
Fábio André da Cunha Almeida · 2012
Consumption of multimedia content, from TV programming, to music downloads through video streaming, has become common in our daily lives. However, the sheer size and diversity of currently available online resources, is turning choice and selection into a very difficult task. Systems that can assist the user in selecting useful information are thus becoming increasingly important. Such systems make use of recommendation engines – this is the main area of this dissertation. This work has been conducted within the context of a research project being led by INESC Porto, to develop a recommender engine for the hospitality market. This engine is to be incorporated in the IPTV entertainment services offered to hotel guests, through which, in addition to watch TV programmes or rent movies, the guests can also access other types of services (e.g. book a tennis court). The initial version of this recommendation engine follows a content-based approach and provides recommendations only for the television service. Three main limitations can be pointed to this initial implementation: 1) it is known that content-based approaches can rapidly become very focussed, providing the user always the same type of recommendations; 2) it only addresses TV content, neglecting other types of resources and services the guest has access to; 3) it is not capable of establishing relationships across both the users and the resources, to increase the performance and enrich the recommendations. The main motivation of this dissertation was to investigate solutions to overcome these limitations. In particular, it aimed at exploring the use of hybrid approaches, by establishing a measure of proximity between guests and thus include in the recommendation list of one guest, items that would have been preferred by other guests with similar profile. By finding similarities between users, it is possible to make completely unexpected recommendations to them – this is the main advantage of collaborative approaches. Given that both content-based (CB) and collaborative filtering (CF) approaches have individually several limitations, an hybrid approach has the potential of delivering enhanced results by exploring the best of both. This goal was pursued by initially performing a study, comparing the performances of different hybrid approaches when applied to the same type of resources as the ones present in our system. This procedure was important, as it has already been proved that the performance of different approaches can significantly vary depending on the domain they are applied to. Supported by this initial study, a novel hybrid approach based on the kNN algorithm was proposed: an improved Pearson Correlation method for the user similarity weight computation was used to select the user’s neighbourhood, which was then used to generate a collaborative “Predicted Profile”. This profile is then used to generate recommendations applying the CB method that had been previously implemented (Naive Bayes). This novel method proposed in this dissertation adopts a different approach to address the sparsity problem, since there isn’t a user-item rating matrix used in recommendations generation. This can be considered very important, especially when dealing with a large quantity and diversity of resources.